Inverse Design of Tunable Lowpass Microstrip Filters Based on Generative Adversarial Network and Transfer Learning
Yuwei Zhang, Jinping Xu, Shiqi Jiang
Abstract
In this paper, a novel inverse design model, constructed using a generative adversarial network (GAN), is proposed for a varactor-based microstrip tunable lowpass filter (TLPF). Transfer learning algorithm is utilized to simplify the training process of the constructed model. The planar resonant circuits of the TLPFs are represented as pixelated patterns and tunable characteristics are achieved by the tunable capacitance of the varactors. The GAN-based model incorporates multiple simulators that are constructed using convolutional neural networks (CNNs) to predict the outcomes under various tuning conditions. In this way, the relationship between these patterns and the corresponding EM behaviors is established under different tuning voltages applied to the varactors. To significantly reduce the training time and the number of samples required by the extra simulators, an effective solution based on transfer learning is proposed by expanding a single CNN of a well-trained simulator to a series of CNNs of similar simulators with a common architecture. Utilizing the inverse design model, four TLPFs with different customized tuning ranges within 5 to 10 GHz are designed with an efficient process. The training of the model requires approximately 201 min, while the inference time for each design is 12.5 min. The simulation results of the designed TLPFs are in good agreement with the customized ones. Circuit prototypes combining these TLPFs and two additional rejection cells were fabricated and the measurement results show that a 17 dB rejection level is achieved over a wide stopband up to 28 GHz.